Hidden model Obsidian 3 tops FrontierCode 2.0 at 46.7% — but costs 3.6x Opus 5.5
teortaxesTex · x · 2026-10-28
- Opus 5.5 [high] scores 34.5% at $2.42/rollout; [max] hits 45.5% at $12.82.
- Obsidian 3 [high], a hidden model in FrontierCode 2.0, scores 46.7% at $8.62 — about 3.6x Opus 5.5's price at the same effort level. Pushing to [max] ($47.44/rollout) actually drops it back to 45.5%, with no further gains.
- Clues: it runs on the claude-code harness, which normally only six Claude models use, fueling speculation it's an unreleased Anthropic model (possibly "Fable 5.5").
- Author asks: is the FrontierCode harness broken, or is Fable 5.5 a flop?
The attached screenshots show per-effort-level JSON metrics (correct rate, tokens, cost, flaggedrate) for both models.
More from Models
- IBM's STAIR retriever cuts hallucination 65x vs fine-tuned RAG, hits 82.6% Recall@1 — Promptmethus · 2026-10-28
- One prompt burned Gemini 4's entire 5-hour quota; Mammouth AI used just 4% — fazerdino · 2026-10-28
- Why OpenAI and Anthropic keep swapping places on reasoning evals — teortaxesTex · 2026-10-28
- Reddit debates the (unconfirmed) Gemini Argo Ultra usage limits — Qubit99 · 2026-10-28
- OpenBMB open-sources MiniCPM-OCR: 1B SOTA OCR plus a 2B accuracy-checker model — alifcoder · 2026-10-28
- Opus 5.5 burns 80% of user's OpenAI weekly quota in 2 hours on Terra — mgostIH · 2026-10-28